Filling and Disentanglement: Toward Low- and High-Order Parallel Single-Domain Generalization for SAR Ship Detection

Abstract

Unsupervised domain adaptation (UDA) has shown promising results in ship detection tasks for synthetic aperture radar (SAR) images under distribution shifts. However, its effectiveness is contingent upon the availability of unlabeled target data to mitigate domain discrepancies during training, which poses a challenge in real-world scenarios, as acquiring SAR data from the target domain can be time-consuming and resource-intensive. Moreover, it is impractical to retrain the UDA model whenever the target domain changes. To address this, we propose a domain-generalized SAR ship detection framework designed to train a model solely on a single source domain, enabling direct and efficient application to different target domains without requiring target domain data during training. We introduce a novel model that enhances the generalization of SAR ship detection at two levels. On the one hand, due to the limited space of feature distributions in a single source domain, a simple and effective interpolation method named low-order latent space filling module (LoFi) based on feature normalization is proposed. This allows the framework to simulate images captured by different devices, enabling the feature extractor to learn generalized feature representations. On the other hand, to better train a more cross-domain stable detector, we propose a high-order instance disentanglement module (HoDi) based on contrastive learning, which couples the original feature decomposition into task-relevant and task-irrelevant features through instance-level contrastive loss and an entropy loss as an additional constraint. Experiments conducted on four distinct SAR ship datasets obtained from different satellites validate the effectiveness of the proposed model.